Using Gradient Descent Optimization for Acoustics Training from Heterogeneous Data

In this paper, we study the use of heterogeneous data for training of acoustic models. In initial experiments, a significant drop of accuracy has been observed on in-domain test set if the data was added without any regularization. A solution is proposed by getting control over the training data by optimization of the weights of different data-sets. The final models shows good performance on all various tests linked to various speaking styles. Furthermore, we used this approach to increase the performance over just the main test set. We obtained 0.3% absolute improvement on basic system and 0.4% on HLDA system although the size of the heterogeneous data set was quite small.